How Answer Engines Choose Their Sources
Answer engines run searches, read a shortlist of pages, and build one answer from the clearest passages they find. They don't reward the best business. They reward the best evidence. This chapter explains the machinery, because every decision in this book follows from it.
Somewhere this week, a buyer who would have been perfect for your business asked an AI assistant a question you could have answered better than anyone. The assistant replied in about four seconds. It named two or three businesses, quoted a line or two from their websites, and the buyer moved on. There was no results page. There were no ten blue links. There was an answer, and you were either in it or you weren't.
This book is about being in it. Before any of the tactics make sense, though, you need to know how the machine on the other side works. Not at a computer science level. At the level a business owner needs: what the engine does with a question, where it looks, and what makes it pick one source over another.
What an answer engine is
An answer engine is any system that responds to a question with a composed answer instead of a list of links. The ones that matter commercially right now are ChatGPT, Google's AI Overviews and AI Mode, Perplexity, Gemini, and Microsoft Copilot. They differ in detail, but they share one architecture, and that shared architecture is the useful thing to understand.
None of them knows about your business in any permanent sense. A model's training data is a snapshot that starts ageing the moment it's created, so for anything current, local, or commercial, the engine doesn't rely on memory. It goes and looks.
The four steps between a question and an answer
When someone asks a commercial question, the engine does four things in a few seconds.
First, it rewrites the question. One query becomes several. "Best physio near me for a running injury" might fan out into searches for sports physiotherapy clinics in the suburb, running injury treatment options, and reviews of clinics it has seen mentioned before. This step matters more than it looks: your website is not competing on the question the buyer typed. It's competing on the handful of searches the engine derived from it.
Second, it retrieves. Those searches run against ordinary web indices, mostly Google's and Bing's. What comes back is a shortlist of pages, typically somewhere between five and thirty. If your site doesn't surface at this step, nothing else in this book can help for that question. This is why you'll hear us say that SEO isn't dead. It's the entry ticket. It's just no longer the prize.
Third, it reads. The engine pulls the shortlisted pages and looks for passages that answer the question directly. Not pages. Passages. A specific paragraph, a table, a clearly labelled price, a sentence that says the thing plainly. A page can rank well and still contribute nothing to the answer, because nothing on it can be lifted cleanly.
Fourth, it composes. The engine writes one answer from the best passages it found, attributes some of them, and names the businesses its evidence supports naming. Then it stops. The buyer reads a single answer built from perhaps five sources, and for that buyer, in that moment, those five sources are the market.
What gets a source chosen
Sitting behind those four steps are four selection pressures, and they're more specific than "good content".
The first is retrievability. The engine can only lift from what its searches return, so the unglamorous fundamentals hold: indexed pages, sound technical SEO, rankings within reach for the questions your buyers ask. The second is liftability. The engine wants passages it can extract without repair work: a question answered in the first sentence under a heading, a price stated as a number, a claim that stands on its own without needing the three paragraphs above it. The third is attribution. The engine needs to know who is speaking. A page that clearly belongs to a named business, with a named author, consistent details, and a connected footprint across the web, is a safer thing to cite than an anonymous one. The fourth is corroboration. Engines are cautious by design. A claim that appears in one place is a claim; the same claim appearing across your site, a directory, a review platform, and an industry publication is a fact. Sources that agree with the rest of the web get quoted. Outliers get skipped.
Here is the opinion this whole book is built on: answer engines do not find the best business. They find the best-evidenced business. That's uncomfortable if you're excellent and under-documented, which describes most of the established Australian businesses we work with. It's also the entire opportunity, because evidence is buildable, and most of your competitors aren't building it either.
Where this leaves you
Everything that follows works backwards from this machinery. Part II shows you how to audit where you stand today: which questions matter for your business, what the engines say when they're asked, and whose evidence is being used. Part III is the work itself, and it maps one to one onto the four selection pressures you just read. Part IV turns it into an operation with a monthly cadence, because the engines re-retrieve constantly, and visibility is a position you hold, not a badge you win.
One more thing before the next chapter. If you're waiting for this to settle down before acting, it won't. The interfaces will keep changing and the market share between engines will keep moving. But the architecture, retrieve then read then compose, has been stable across every major engine since they launched, and the selection pressures are stable with it. That's what you build against.